Unmanned aerial vehicle pipeline inspection control method based on model prediction contour algorithm

Abstract

The present application relates to a kind of unmanned aerial vehicle pipeline inspection control method based on model prediction contour algorithm, to be based on the unmanned aerial vehicle platform of laser radar and monocular camera, to the three-dimensional pipeline point cloud with semantic label as input, to the joint state of unmanned aerial vehicle body position, speed, path progress and cross section offset, to the unified optimization goal of contour error, lag error, forward speed, control smoothness, two-dimensional offset and obstacle avoidance soft constraint, constructs closed loop inspection control system suitable for complex industrial pipeline scene.

Description

Technical Field

[0001] This invention relates to unmanned aerial vehicle (UAV) control technology, specifically to a UAV pipeline inspection control method based on a model prediction profile algorithm. Background Technology

[0002] In modern large-scale steel smelting and rolling enterprises, a complex network of external media transport pipelines is densely distributed between blast furnaces, converters, continuous casting machines, and various hot and cold rolling production lines. These external pipelines are responsible for transporting high-temperature steam, gas, oxygen, cooling water, and various chemical additives, serving as the physical arteries that maintain continuous steel plant production. Due to the vast area of ​​steel plant sites, these transport pipelines are constantly exposed to high temperatures, high dust levels, strong corrosion, and industrial exhaust gases. Traditional external pipeline inspections heavily rely on manual climbing of scaffolding or the use of specialized aerial work platforms. This not only consumes enormous manpower and resources and is extremely inefficient, but also poses significant safety hazards in the complex industrial environment, which is fraught with the risk of gas leaks and strong magnetic field interference.

[0003] In recent years, drones have gradually entered various industries, especially due to their high mobility and flexibility. They are also being introduced into factories for external inspection of overhead pipelines, thermal imaging temperature measurement, and weld flaw detection. However, the external pipeline network of steel plants is not isolated; it is often surrounded by various supporting trusses, valve protrusions, walkway flanges, and crisscrossing cable trays. In the low-altitude environment where GPS signals are severely affected by multipath effects or even blocked due to the dense steel structures, requiring drones to maintain a fixed safe operating distance close to the outer wall of the pipeline and perform high-precision autonomous flight and target tracking beyond visual range poses extremely demanding theoretical challenges to the underlying flight control algorithms.

[0004] In existing technologies, the main control methods for UAV pipeline inspection include manual inspection, PID control, MPC control, artificial potential field control, and visual tracking. Manual inspection requires skilled operators, and all control is based on manual operation using personnel and camera images, which is not only complex but also has a high error rate. PID, MPC, and artificial potential field plus vector field control methods are more inclined towards time-domain tracking, and cannot closely follow the pipeline perimeter for pipeline inspection, leading to unstable control and oscillations in complex environments. Visual tracking inspection is suitable for single-pipeline tracking and simple environments, but in complex environments with multiple pipelines, control instability is likely to occur. To achieve accurate and close-fitting inspection in complex environments, there is an urgent need for a control method that can respond quickly and safely for tracking and inspection. Summary of the Invention

[0005] The purpose of this invention is to provide a UAV pipeline inspection control method based on a model prediction profile algorithm that can simultaneously achieve path continuity, control stability, observation effectiveness, and obstacle avoidance safety in industrial pipeline inspection.

[0006] The present invention adopts the following technical solution:

[0007] A UAV pipeline inspection control method based on model prediction contour algorithm includes the following steps:

[0008] (1) By processing the input pipeline point cloud data with semantic labels, coordinate units are unified and point sets are classified;

[0009] (2) Cluster the endpoints, elbows, and valve points to obtain the connector center and POI candidates; construct the straight pipe center axis and establish a topology graph with straight pipe sections and connectors as nodes;

[0010] (3) Generate discrete inspection track points based on the topology map, and generate a safety inspection trajectory by combining the offset mode; perform Gaussian smoothing and PCHIP interpolation on the discrete trajectory points to obtain the path function. and its derivative and calculate the curvature;

[0011] (4) Construct the Parallel Transport section base along the path Calculate the nearest distance and corresponding parameters to all POIs, filter the valid POIs for the paths, and sort them.

[0012] (5) Establish an extended state model; in each sampling period, based on the current state, curvature, reciprocating manifold area, POI state and obstacle distance, solve the quadratic programming online and output the acceleration, path advancement speed and cross-sectional offset change rate;

[0013] (6) When the UAV approaches the POI, it enters the approach and hovering inspection mode and performs multi-view photography; when all inspection tasks are completed, the UAV reaches the destination and performs the ascent and hovering end action.

[0014] Furthermore, in step (1), the single-point data... Perform a unified unit conversion to obtain the set of straight pipe points and the set of non-straight pipe points; This represents the three-dimensional coordinates of the pipe relative to the global coordinate system. The label represents the normal vector value for the regional pipeline cluster after subsequent calculations. i For semantic tags.

[0015] Furthermore, in step (2), the specific method for establishing the topology graph is as follows:

[0016] (a) For any straight pipe segment, establish its axis start and end points by extracting the main axis. And construct the axis:

[0017]

[0018] Where τ is a random point parameter of the pipe segment axis; when τ is 0, it is the starting point, and when it is 1, it is the ending point.

[0019] (b) Establish a diagram structure based on each straight pipe section and each connector. ;

[0020] If the endpoint of the straight pipe section and the center of a certain connector satisfy the following... , Therefore, edges are added between the corresponding nodes. This is the topology connectivity threshold; The center of a certain connector; if there are pipe segment nodes in the diagram with a degree less than the requirement, topology repair is performed according to the minimum distance criterion;

[0021] If multiple connected components exist, then for any two components Select ;in This represents the spatial location of a graph node; adding connecting edges to this node yields a globally traversable topology.

[0022] Furthermore, in step (3), the formula for calculating the curvature is:

[0023]

[0024] in, r(θ) is the path function;

[0025] set up The final path function is , , To each , , Obtained by performing PCHIP interpolation; For discrete points of the trajectory, This is the cumulative arc length.

[0026] Furthermore, in step (4), the Parallel Transport section base The construction method is as follows: along the discrete parameter sequence Calculate the tangent vector in the direction At the initial point, if the unit vector of the pipeline is... If the reference axis is not parallel to the world vertical axis, select the world vertical axis as the initial reference; otherwise, select the world horizontal axis as the reference. Let the initial reference axis be... The initial normal basis is defined as , For subsequent discrete points, if the change in adjacent tangential directions is significant, propagation is performed according to the rotation formula; if the change is minor, the normal basis from the previous step is directly maintained; finally, for any continuous parameter... The result is obtained by linear interpolation and normalization of the nearest discrete basis vectors. .

[0027] Furthermore, in step (4), the method for filtering and sorting valid POIs along the path is as follows: all connector centroids and external input POIs are uniformly represented as... , Indicates the POI location. For POI normal direction, Define the semantic labels for POIs; calculate the minimum distance from each POI to the path: Only when satisfied Only then is the POI considered a valid POI for the path, and its corresponding path parameter is recorded as... , The path association threshold; all valid POIs are set according to... After sorting, a sequence of key points for scheduling inspection tasks is formed.

[0028] Furthermore, in step (5), the extended state model is a discrete linear extended state model, and the state and control quantity at the k-th sampling time are respectively expressed as:

[0029] ,

[0030] Discrete systems ,

[0031] , , and These are the three-dimensional and two-dimensional identity matrices, respectively.

[0032] The update process corresponding to the extended state model is as follows: .

[0033] Furthermore, in step (6), the error cost of solving the quadratic programming problem at step j is:

[0034]

[0035] in, for The transpose of ;

[0036] For a comprehensive weighted matrix,

[0037] in, For the contour error weights of a single step, It is the weight of the hysteresis error; and It is the projection matrix of contour error and hysteresis error; the contour error is the projection of the difference between the actual UAV position and the path point on the prior path line onto the direction of the vertical pipe midline, and the hysteresis error is the projection of the corresponding difference onto the direction of the parallel pipe midline.

[0038] For the first The error vector of the step.

[0039] Among them, the error state coefficient matrix x j The remaining parameters after removing state variables from the state at time j. ; For path parameters, path points tangent vector of path point The path point corresponds to the cross section base. .

[0040] Furthermore, in step (6), the cost function of the extended state model is:

[0041]

[0042] in, To track error terms, To control input penalty items, For fluid-level control smoothing terms, For reference, track progress items, For cross-sectional regression terms, This is the relaxation term of the mixed-line quadratic form.

[0043] Furthermore, in step (7), after entering the approximation mode, the mode velocity factor is set to...

[0044]

[0045] To approximate the distance threshold for POI, Minimum speed factor For the progress of the i-th POI Compared with the current progress The distance between them;

[0046] The arrival criterion is satisfied or and Then it switches to hover inspection mode; among which, The progress difference threshold, For POI target points, Reaching the distance threshold, Speed ​​threshold.

[0047] The beneficial effects of this invention are as follows: This invention integrates inspection object understanding and flight control through semantic tag point cloud, topology map reconstruction, and automatic POI generation; it achieves a unified expression of advancing along the center curve and flexibly inspecting within the normal section through joint modeling of path progress variables and cross-section offset variables; it significantly improves path fit and forward stability in complex 3D pipeline scenarios by decoupling contour error and hysteresis error; it improves feasibility and safety in locally confined environments by directly introducing obstacle avoidance soft constraints into quadratic programming and combining them with KD trees to correct the flight trajectory in the relative obstacle normal direction in real time based on nearby obstacle determination; and it provides strong engineering robustness and mission integrity through curvature adaptation, reciprocating manifold area deceleration, dynamic POI selection, and post-inspection speed recovery. Attached Figure Description

[0048] Figure 1 This is a bird's-eye view of a pipeline in a simple pipeline inspection simulation scenario.

[0049] Figure 2 for Figure 1 The performance evaluation of the controller in the simple pipeline inspection simulation is shown.

[0050] Figure 2 In the diagram, (A) is the contour error result diagram, (B) is the hysteresis error result diagram, (C) is the 2D cross-section offset result diagram, and (D) is the safety distance monitoring result diagram.

[0051] Figure 3 A bird's-eye view of a pipeline in a complex pipeline inspection simulation scenario.

[0052] Figure 4 for Figure 3 Performance evaluation of the controller in the simulation of complex pipeline inspection shown.

[0053] Figure 4 In the diagram, (A) is the contour error result diagram, (B) is the hysteresis error result diagram, (C) is the 2D cross-section offset result diagram, and (D) is the safety distance monitoring result diagram.

[0054] Figure 5 The graph shows a comparison of the computational complexity of KD-tree and brute-force search.

[0055] Figure 5 In the table, (A) represents a comparison of query efficiency (speedup = 3.0x), and (B) represents a comparison of query times per iteration.

[0056] Figure 6 Arrival time of POIs for simple pipelines and execution of shooting statistics (8 points, 40 photos).

[0057] Figure 6 In the table, (A) represents the POI access time, and (B) represents the photo statistics results for each POI.

[0058] Figure 7 Arrival time of POIs for complex pipelines and execution of shooting statistics (20 points, 100 photos).

[0059] Figure 7 In the table, (A) represents the POI access time, and (B) represents the photo statistics results for each POI.

[0060] Figure 8 This is a diagram illustrating the effect of adaptive speed adjustment based on the curvature of a simple pipe and the POI.

[0061] Figure 9 This diagram illustrates the effect of adaptive speed adjustment for complex pipeline curvature and POI. Detailed Implementation

[0062] This invention constructs a UAV pipeline inspection system consisting of a UAV flight platform, an airborne perception module, a semantic point cloud processing module, a pipeline topology modeling module, a path generation module, a Parallel Transport section base module, an MPCC optimization control module, a POI inspection state machine module, and a safety protection module.

[0063] From the overall methodology, this invention can be summarized as follows: Inputting semantically tagged pipeline point cloud data via an airborne system or offline, completing coordinate unit unification and point set classification; clustering endpoint, elbow, and valve point sets to obtain connector centers and POI candidates; performing connector neighborhood trimming, clustering segmentation, and principal axis extraction on straight pipe point sets to construct the straight pipe center axis; establishing a topology graph using straight pipe segments and connectors as nodes, and performing adjacency and connectivity repair; generating discrete inspection track points based on the topology graph, and generating a safety inspection trajectory by combining the offset mode; performing Gaussian smoothing and PCHIP interpolation on the discrete trajectory points to obtain the path function. and its derivative; constructing the Parallel Transport section basis along the path The process involves calculating the nearest distance and corresponding parameters to all points of interest (POIs) along the path, filtering and sorting valid POIs, establishing an extended state model, error projection matrix, and soft constraints on obstacles, solving a quadratic programming problem online in each sampling period based on the current state, curvature, reciprocating manifold area, POI state, and obstacle distance, and outputting acceleration, path advancement speed, and cross-sectional offset rate of change. When approaching a POI, the drone enters an approximation and hovering inspection mode, performing multi-view photography. After all inspection tasks are completed, the drone reaches its destination and performs a hovering ascent to end the inspection. Thus, this invention achieves a complete closed-loop technology chain from semantic point cloud understanding, inspection topology construction, path parameterization to real-time model prediction contour control.

[0064] (a) Mechanical structure

[0065] In terms of mechanical structure, the UAV platform adopts a commercially available multi-rotor aircraft, with a laser camera, a monocular camera, an onboard inertial measurement unit, a flight control computing unit, an onboard computing unit, and a wireless communication module. The laser camera is used to obtain 3D point clouds of the environment, the monocular camera is used to complete POI visual observation and image acquisition, the onboard computing unit is used to complete point cloud reading, path generation, model prediction optimization, and state switching control, and the flight control computing unit is used to perform position, velocity, and attitude inner-loop stabilization control.

[0066] (ii) Prior path generation based on pipeline point cloud

[0067] (1) Acquisition of point cloud data

[0068] Raw point cloud data, i.e., 3D location information data The points are obtained in two ways: (1) by scanning the actual pipeline using a handheld laser scanner (with built-in IMU) or a drone equipped with a laser radar and IMU. (2) by drawing a 3D model of the pipeline using SolidWorks and then outputting the overall point cloud in binary format using Python code.

[0069] Obtaining complete point cloud data: After adding the corresponding normal vectors calculated by Python to the obtained raw point cloud data, six-dimensional data is obtained. Building upon this, semantic annotations of pipeline scenarios and components are added (using CloudCompare software) to ultimately form complete point cloud data. .

[0070] The semantics of the complete point cloud need to be manually labeled in the early stage. Subsequently, the pipe components such as valves, elbows, straight pipes and endpoints are automatically segmented by the pipe component segmentation model pre-trained by PointNet, and the complete point cloud data is directly output. That is, the input point cloud data is used as the basis for subsequent calls.

[0071] The output method of POI centroid data is as follows: The density of the input point cloud data in the pipeline is determined by Python code (clusters with the same label and adjacent different clusters). The centroid of the POI region is geometrically calculated and output. Its data type is (component_id region ID, label label category, num_points number of points in this region, centroid_x, centroid_y, centroid_z centroid 3D coordinates, normal_x, normal_y, normal_z corresponding 3D normal vector of this centroid). All subsequent POI data operations are derived from this.

[0072] In simple terms, the raw point cloud data of the pipeline is collected or mapped by on-site radar and is three-dimensional data (position information). After initial processing, it becomes six-dimensional data (position information + normal vector), and further processing results in seven-dimensional data (position + normal vector + label). Points of Interest (POI) regions are extracted based on the label information. All points in the pipeline point cloud, excluding straight pipes with label=0, are designated as POI regions. After geometric fine-grained clustering, regions are classified as Region 1, Region 2, etc. The same label can also be divided into multiple POI regions (e.g., pipelines). Each POI region is sorted according to the pipeline centerline direction, and the POI centroid data is extracted. POI centroid data can be obtained by inputting pipeline point cloud data, or by inputting externally calculated and saved data along with the pipeline point cloud data. The algorithm will determine if existing POI centroid data is currently input; if not, it will automatically determine and calculate the extracted POI centroid data to prepare for subsequent operations.

[0073] (2) Point cloud preprocessing and pipeline structure analysis

[0074] The inspection system inputs pipeline point cloud data with semantic tags, and its single-point data format is as follows:

[0075]

[0076] in, As a synonym for the initial input point cloud (an input point cloud that has already been adjusted and has normal vectors and labels), This represents the three-dimensional coordinates of the pipe relative to the global coordinate system. () represents the normal vector value for the regional pipeline cluster after subsequent calculations. Indicates a straight pipe. This indicates the two ends of the pipeline, i.e., the start and end points. Indicates a bend. Indicates a valve.

[0077] The inspection system first performs a unified conversion of coordinate units. If the point cloud source unit is millimeters, then it is adjusted according to the scaling factor. The transformation is performed to obtain the three-dimensional coordinate set of the pipeline. If the original data is already in meters, then take... This process ensures that the point cloud scale is consistent with the flight control physical parameters, which is a prerequisite for subsequent radius estimation, safety distance setting, and velocity constraints.

[0078] For the straight pipe point set For points with a semantic label of 0, first estimate the average distance between points. ,in, Point The nearest neighbor points in the local neighborhood are used to estimate the pipe radius within the local neighborhood. Random sampling of the center point is then performed. Surrounding neighborhood point set We construct its decentralized coordinate matrix and obtain the local principal direction and two cross-sectional directions through principal component analysis.

[0079] If the widths of the local section in the two secondary principal directions are respectively:

[0080]

[0081] in, This represents the two secondary principal directions obtained through principal component analysis. This represents the point set of this region. The transpose of each point.

[0082] Therefore, projection calculations are performed in two directions, and the local radius is estimated as follows: The pipe radius estimate is obtained by taking the median value of multiple sample radii. This radius is used not only to define the safety inspection offset distance, but also for connector cutting, topology connection threshold, and POI validity screening threshold.

[0083] For non-straight pipe point set Points with non-zero semantic labels are clustered using density clustering to form connector clusters. Let the j-th cluster after clustering be...

[0084]

[0085]

[0086] Cluster Center Used to construct graph nodes and POIs. This is the distance from the farthest point in the same cluster to the cluster center (i.e., the cluster scale or the radius of the farthest point), used to define the local trimming radius of the connector.

[0087] Next, the radius around each connector (i.e., the radius of the farthest point) ) plus buffer radius The straight pipe point inside is from After removing the middle section, the set of remaining straight pipe points is obtained. Buffer radius Three times the average point cloud spacing is used to ensure complete removal. This avoids interference from complex geometry near elbows, valves, and ends in the extraction of the straight pipe's center axis. Subsequently... Spatial clustering was performed to obtain several candidate straight pipe segments. .

[0088] For any given pipe segment, its starting and ending points are established through spindle extraction. And construct the axis:

[0089]

[0090] The diagram structure is established based on each straight pipe section and each connector. If the endpoint of a straight pipe section and the center of a certain connector satisfy the following:

[0091]

[0092] Then add edges between the corresponding nodes. Here This is the topology connectivity threshold, typically a linear combination of pipe radius and node spacing. If there are pipe segment nodes in the graph with a degree less than the required value, topology repair is performed according to the minimum distance criterion. If multiple connected components exist, then for any two components... Select:

[0093]

[0094] in, This indicates the spatial location of a graph node. Then, connecting edges are added to the node pair to obtain a globally traversable topology.

[0095] This process corresponds to the preparatory work before path generation: connector clustering, straight pipe cutting, axis extraction, and topology repair stages.

[0096] (3) POI extraction and association, inspection path generation

[0097] During the path generation phase, the starting and ending points of the traversal are determined from the topology graph. If an endpoint node with label 1 exists, one of them is selected as the starting point, and the other endpoint with the greatest Euclidean distance from it is selected as the target endpoint. When the shortest path can be found, it is... This is a sequence of nodes traversed on the graph; if the shortest path solution fails, a depth-first search is used to obtain an alternative sequence. For each straight pipe segment node in the path, its axis is extracted and its direction is adaptively determined according to the position of the previous node to ensure the consistency of the trajectory points. For each connector node (i.e., point of interest), its center point is used as the key bridging point, and a POI point is generated using the local normal at that center point. If the system operates in centerline following mode, the discrete trajectory points on the straight pipe segment are the axis sampling points; if the system operates in offset trajectory mode, an offset vector is constructed for each straight pipe segment.

[0098] Let the direction vector of this segment be:

[0099]

[0100] Take the reference upper vector in world coordinates ,like and Switching the reference axis when the axis is nearly parallel avoids degradation. Then, the auxiliary normal is calculated. Reconstruct offset direction Therefore, the straight pipe section offset trajectory is set as follows:

[0101]

[0102] For safe inspection distance.

[0103] To prevent adjacent segments from flipping over in their offset directions, if the offset direction of the current segment satisfies the same condition as the offset direction of the previous segment... Then Inverting the path line is a design feature to prevent it from flipping, significantly improving the consistency of trajectory connections between segments. After concatenating all discrete trajectory points, Gaussian smoothing and PCHIP interpolation are used for spatial path parameterization.

[0104] Let the discrete points of the smoothed trajectory be... The cumulative arc length is Then the path function .

[0105] Depend on , , To each , , Obtained by performing PCHIP interpolation. Its first and second derivatives are as follows:

[0106]

[0107] The first derivative is used for tangential and velocity mapping, and the second derivative is used for curvature calculation. The curvature calculation formula is:

[0108]

[0109] This curvature reflects the degree of local bending of the path and is an important basis for subsequent adaptive control weights and stable control of the reciprocating manifold area.

[0110] During the construction of the Parallel Transport section basis, this invention follows the discrete parameter sequence. Calculate the tangent vector in the direction At the initial point, if the unit vector of the pipeline is... If the reference axis is not parallel to the world vertical axis, choose the world vertical axis as the initial reference; otherwise, choose the world horizontal axis as the reference. Let the initial reference axis be... The initial normal basis is defined as , For subsequent discrete points, if the adjacent tangential changes significantly, propagation is performed according to the rotation formula; if the change is small, the normal basis from the previous step is directly maintained, thus preventing the UAV from rotating around the pipe. Finally, for any continuous parameter... The result is obtained by linear interpolation and normalization of the nearest discrete basis vectors. This establishes an orthogonal local cross-sectional basis for each point on the path. This defined cross-sectional basis has the advantages of being continuous and not easily flipped, supporting the two-dimensional offset control and POI surround-view photography described later in this invention.

[0111] During the POI generation and filtering phase, all connector centroids and external input POIs are uniformly represented as... , Indicates the POI location. For POI normal direction, This refers to the semantic labels for Points of Interest (POIs). The output format after processing the POI data is (cluster block ID, semantic label, number of point clouds, 3D coordinates of the centroid, and 3D normal vector of the centroid). 1×9 In the actual input and utilization stage, the data content is automatically extracted according to the column number, thus obtaining a representation in the same format as the original pipeline point cloud data.

[0112] To determine whether a POI is within the reach of the planned inspection path, the minimum distance from each POI to the path is calculated: Only when satisfied Only then is the POI considered a valid POI for the path, and its corresponding path parameter is recorded as... , This is the path association threshold, taken as an integer multiple of the pipeline radius. All valid POIs are configured according to... After sorting, a key point sequence for inspection task scheduling is formed. This invention, by directly extracting the centroids of endpoints, bends, and valves based on semantic tags, better ensures consistency between the inspection target and the engineering semantics compared to simple geometric feature extraction.

[0113] (III) Improving the optimal control of MPCC

[0114] In terms of control modeling, this invention employs a discrete linear extended state model. Let the state and control quantity at the k-th sampling time be represented as follows:

[0115] ,

[0116] Discrete systems .

[0117]

[0118]

[0119] and These are the three-dimensional and two-dimensional identity matrices, respectively.

[0120] The corresponding state update process for this model is as follows:

[0121]

[0122] Its advantages are twofold: firstly, it maintains the linear structure of the prediction model, facilitating real-time quadratic programming solutions; secondly, it explicitly incorporates path progress and cross-section offset into the optimization state. Compared to the original position velocity and path progress θ, it innovatively adds two cross-section offset variables, namely the distance of the UAV from the centerline in the cross-section plane perpendicular to the path. This overcomes the shortcomings of the original MPCC, which only tracks the centerline r(θ), such as inflexible obstacle avoidance, inconvenient POI inspection, and oscillation when curvature changes. It enables the UAV to follow the regression prior path normally when there are no POIs or obstacles, and constrains the UAV to flexibly avoid obstacles in the vertical direction of travel when there are obstacles, thereby enhancing the control capability.

[0123] Within a finite prediction domain N, the future states and controls are stacked as follows:

[0124]

[0125] Then there is , This is the state prediction matrix. To control the cumulative matrix, based on this expression, the path error, velocity constraint, offset constraint, and obstacle avoidance constraint can be uniformly written as a quadratic programming problem with respect to U. For the linearized reference point at step j in the prediction time domain, let its path parameters be... Its path points are And the tangent vector is .

[0126] Corresponding cross-section base is The offset of the drone in both directions at this cross section is The predicted path parameters are The predicted location of the drone is Then the linearized geometric mapping is written as:

[0127] .

[0128] To further facilitate matrix transformation, define , Therefore, the error vector at step j can be written as: The error is projected onto the tangential and normal directions respectively for subsequent weight control, in order to better adapt to the inspection task. Define the contour projection matrix. Define the hysteresis projection matrix The overall weighted matrix is ​​then... ,in, For the contour error weights of a single step, It is the weight of the lag error; P c and P l This is the projection matrix of contour error and hysteresis error; the contour error is the projection of the difference between the actual UAV position and the path point on the prior path line onto the direction perpendicular to the pipe midline, and the hysteresis error is the projection of the corresponding difference onto the direction parallel to the pipe midline; therefore, the error cost at step j is... This formula is one of the key innovations of this invention. Because Only the error component perpendicular to the path tangent is retained, corresponding to the contour error; Only the error component along the tangential direction of the path is retained, corresponding to the hysteresis error. Compared to directly penalizing the three-dimensional error with a unified Euclidean norm, this structure is more in line with the concept of contour control, enabling the system to flexibly handle the propulsion error along the path direction while ensuring conformity to the pipe shape.

[0129] Highly Coupled Comprehensive Cost Function Design: The core control mechanism of this invention is embodied in a globally coupled comprehensive quadratic cost function. This function breaks through the limitations of traditional single error penalty and consists of six main parts: spatial tracking term, control input amplitude term, dynamic smoothing term, inspection progress term, cross-sectional regression term, and obstacle avoidance relaxation term.

[0130] .

[0131] Tracking error term: .in, The deviation between the predicted position of the UAV and the reference path point in step k is given. , These are the contour error weights and hysteresis error weights, respectively. , These are the contour error projection matrices, respectively. The system assigns a constant penalty with extremely high weights to the contour error, forcing the UAV to closely fit the expected cross section; while the hysteresis weights are dynamically adaptive.

[0132] Control input penalty items: .in, , , These are the UAV three-axis acceleration input, path progress evolution speed control input, and pipeline cross-section offset speed control input for step k, respectively. , , These are acceleration penalty weights, path progression speed penalty weights, and cross-section offset speed smoothing penalty weights. This applies secondary penalties to the three types of control inputs—acceleration, path progression speed, and cross-section offset speed—limiting the magnitude of the control inputs to prevent excessive maneuvers, ensuring smooth and stable flight, and ensuring that the control inputs meet the physical constraints of the UAV, thereby improving the stability of the control system and flight safety.

[0133] Cross-sectional regression term: ,in, , , where represents the cross-sectional offset regression weight, and represents the two-dimensional offset of the UAV within the pipe cross-section at step k. This offset within the pipe cross-section is penalized, driving the UAV back to the center of the pipe cross-section when there is no obstacle avoidance requirement, ensuring the UAV remains in a safe flight zone and reducing the risk of collision with the pipe wall.

[0134] Fluid level control smoothing term: ,in, , These are the jerk penalty weight and the cross-sectional offset velocity change penalty weight, respectively. , These represent the changes in acceleration and the changes in cross-sectional offset velocity between adjacent control steps, respectively. To avoid oscillations caused by large maneuvers, amplitude penalties are applied to the three-dimensional acceleration, the square of the evolution velocity, and the offset velocity. More importantly, a strict incremental penalty mechanism is introduced. Through secondary penalties on the jerk acceleration and the increase in the cross-sectional lateral slip rate, the UAV maintains a fluid and smooth flight attitude within extremely narrow spaces.

[0135] Reference tracking progress items: , , , These are the adaptive weights for progress advancement, the path progress evolution speed at the k-th step, and the target path progress evolution speed, respectively. By simultaneously adding a quadratic penalty term for the evolution speed and a linear guiding term pointing to the desired inspection speed to the objective function, the drone is continuously and stably propelled along the pipeline like a spring.

[0136] Mixed-line quadratic relaxation term: In obstacle avoidance soft constraints, this invention not only applies a squared penalty to the relaxation variable, but also adds a linear penalty term to ensure that the optimizer can provide an extremely smooth repulsive gradient when slightly approaching the obstacle, rather than an abrupt step response.

[0137] (iv) Determining the path advancement speed

[0138] For the path advancement speed reference, this invention adaptively sets it based on the cruise state, POI approach state, post-inspection recovery state, and curvature state.

[0139] Let the path curvature be The velocity factor is The maximum setting value for path advancement is Then the original path advancement reference can be represented as , This is an empirical proportionality coefficient. Considering the effect of curvature, the curvature adaptive factor is defined as... ,in , , These are the lower bound of the adaptive factor, the curvature influence coefficient, and the current path curvature, respectively. Then the adaptive lag weight and the path advancement weight are respectively... ,in, , Set the lag weight and path advancement weight as input.

[0140] During the speed recovery phase after inspection, to prevent the drone from lingering at low speed for an extended period due to localized deceleration, settings can be directly configured. Through this mechanism, the system can advance efficiently in straight pipe areas, or maintain a conservative approach in bends and reciprocating manifold areas.

[0141] (v) Path avoidance

[0142] In terms of obstacle avoidance modeling, this invention constructs a KD-tree for obstacle points and performs K-nearest neighbor queries several steps before predicting the trajectory. Let the j-th predicted position be... The set of nearest neighbor obstacle points obtained by the query is For each candidate obstacle point, calculate the escape direction. If the original normal direction of the point cloud The angle between the direction of escape and the direction of flight is small, which means that the condition is met. ,in If the normal angle threshold is used, then the fusion barrier normal is constructed as follows:

[0143]

[0144] Otherwise, take directly. Then, from the K candidates, the one with the worst safety margin is selected as the main constraint direction for this prediction step. If the distance from this point to the prediction position is... The minimum safety clearance is Then the linear safety constraint is written as ,in , , These are the minimum safety clearance, the deviation distance between the linearized reference position and the obstacle point, and the obstacle avoidance relaxation variable, respectively. The position is determined jointly by the linearized reference position and the obstacle point position. This equation is then incorporated into the quadratic programming inequality constraint set. ,in This indicates that the control sequence and slack variable sequence will be solved jointly. The advantage of doing so is that it addresses both short-term security risks and avoids excessive obstacles and constraints in long-term predictions that would slow down the solution process.

[0145] In addition to obstacle constraints, this invention also sets boundaries for velocity and cross-section offset. Predicted velocity constraints.

[0146] in, The dynamic maximum speed limit can be dynamically scaled based on inspection status, reciprocating manifold area, and safety distance. (Section offset constraint) Offset change rate constraint Path advancement speed constraint Acceleration constraints .in , , , , These are the maximum upper limit of cross-sectional offset, the maximum upper limit of offset velocity, the maximum upper limit of minimum and maximum propulsion velocity, and the maximum upper limit of maximum acceleration.

[0147] With these constraints, the system is able to output smooth, safe, and executable control sequences within the physically realizable range.

[0148] (vi) Solver

[0149] The solver of this invention employs an SQP-style iterative mechanism based on quadratic programming. It first constructs an initial guess using the control sequence from the previous time step or the default propagation speed. Then according to The initial predicted state sequence is obtained. In the s-th iteration, around Linearizing the error term and obstacle constraints, we solve the quadratic programming problem to obtain... If the state prediction change satisfies Then terminate the iteration.

[0150] If the difference between the previous prediction and the current initial prediction is small enough, the number of SQP iterations can be reduced to one step to reduce online computation; if the difference is large, two iterations are maintained to ensure the stability of the solution. This adaptive strategy is based on the idea of ​​dynamically adjusting the number of SQP iterations according to the change in prediction.

[0151] The system calculates the progress points on the current path point, i.e., the reference path, in real time during runtime. , , , The three-dimensional coordinates of the current path point, and in the local window , Within the window radius. Search for the nearest parameter on the path from the current position. When the tracking error is too large and the current location is not within the reciprocating manifold buffer zone, a local progress reset is performed. This mechanism prevents severe mismatch between the UAV and path parameters due to momentary disturbances or outer obstacle avoidance maneuvers, thereby improving the stability of long-distance inspections. Resetting occurs only within a local window, and setting a reset buffer in the reciprocating manifold area prevents path index jumps.

[0152] (vii) State machine control

[0153] Regarding the POI dynamic selection state machine, the system maintains a set of unvisited POIs. Calculate for each unvisited POI and select the one that satisfies And the smallest candidate is used as the current active POI, where A backward window that allows for supplementary inspections. For the progress of the i-th POI Compared with the current progress The distance between them. After entering approximation mode, set the mode velocity factor to...

[0154]

[0155] To approximate the distance threshold for POI, As the minimum speed factor, this expression means that the closer to the POI, the more actively the system will decelerate; the farther away, the normal cruising speed will be maintained. The system switches to hover inspection mode when the arrival criterion meets the following conditions: or and ,in The progress difference threshold, For POI target points, Reaching the distance threshold, Speed ​​threshold.

[0156] (viii) POI Inspection Logic

[0157] In patrol inspection mode, set For each sampling time, the hover time accumulator satisfies:

[0158]

[0159] The photo timer meets the following requirements:

[0160]

[0161]

[0162] in, This refers to the maximum hovering time of the drone. Let be the total number of photos taken. A photo is taken and the viewpoint index is updated when the above formula is satisfied. The circumferential viewpoint of the j-th photo is defined as... To obtain the current position of the target observed in the panoramic view:

[0163]

[0164] in, These are the center coordinates of the i-th POI. The radius of the surrounding shooting area is the distance from the camera to the center of the POI. It is the unit normal vector of the POI, and the orthogonal basis vector of the POI section. It consists of the cross product of the POI normal and the reference vector. Let the visited parameter be... After all the photos are taken, And set a speed recovery timer after the inspection. Forced recovery during the recovery period. and This helps the drone quickly leave the low-speed area near the POI and return to its normal inspection rhythm.

[0165] (ix) Emergency obstacle avoidance

[0166] Regarding safety, this invention includes an emergency obstacle avoidance mode. If the distance from the current machine position to the nearest obstacle meets the requirements... , in , The distance from the drone's current position to the nearest obstacle and the emergency obstacle avoidance distance threshold are respectively used. MPCC optimized control will no longer be invoked (see Part (III)). Instead, the emergency control acceleration will be directly output.

[0167]

[0168] The velocity damping coefficient, For escape gain, The unit direction away from the nearest obstacle. The path advance speed is then set to a small positive value. The cross-sectional offset change rate is set to zero to ensure that the system prioritizes escaping collision risks in the most dangerous local areas. Once the distance to the nearest obstacle returns to a safe range, the system re-enters MPCC optimized control (see Part (III)). This two-layer structure, consisting of an outer emergency safety layer and an inner optimizer, balances rapid response in extreme situations with high-quality control in normal conditions.

[0169] Each formula and parameter in this invention has clear engineering significance. (State variables) Instead of a regular time index, it uses geometric progress along the path to separate "how much progress" from the current position. It is not a distractor, but rather the degree of freedom of inspection on the normal section, which enables the drone to move around the pipeline in a surface form; Used to characterize forward and backward deviations in the direction of advance, its reduction helps to avoid progress lags in bends and branch areas; Used to characterize the degree of lateral deviation from the contour, its reduction helps ensure that the body always fits the pipeline inspection belt; curvature Used to reflect the geometric complexity of the path, its participation in weighted scheduling makes the system more conservative in high curvature regions; slack variables Used to make controlled trade-offs between feasibility and safety, avoiding unsolvable optimization problems in narrow environments; POI parameters This maps discrete key components onto continuous paths, ensuring that the inspection logic and control logic maintain the same parameter system.

[0170] Because these variables and relationships are uniformly incorporated into the same optimization framework, this invention can simultaneously achieve path continuity, control stability, observation effectiveness, and obstacle avoidance safety in industrial pipeline inspection. Therefore, the UAV pipeline inspection control system and method based on model predictive contour control algorithm proposed in this invention is not a simple patchwork of existing point cloud processing, path planning, visual inspection, and MPCC control, but rather a unified technical solution formed around the core idea of ​​semantically driven pipeline geometric representation, path contour error decoupling, normal section offset optimization, real-time solution of obstacle soft constraints, and closed-loop execution of POI tasks. This solution has strong engineering deployability and control theory integrity, and can be directly applied to autonomous UAV inspection tasks in various industrial pipeline scenarios.

[0171] (x) Autonomous Inspection Process

[0172] Extracting the prior trajectory line: using the given pipe point cloud (pipe 3D coordinates, pipe 3D normal vectors). 1×6 The data is formatted as follows: the pipe component segmentation model pre-trained in PointNet automatically segments each component of the pipe point cloud and outputs data with the original data and semantic labels, in the format of (pipe 3D coordinates, pipe 3D normal vectors, semantic label). 1×7 For now, we pre-define label: 0 represents a long, straight pipe, 1 represents both ends of the pipe, 2 represents an elbow, and 3 represents a valve. This invention uses this for testing, but can also be further segmented based on specific areas of interest. We consider all labels except 0 as areas of interest for pipe inspection. Based on the point cloud clusters containing these labels, we segment the disorganized actual pipe component areas. After processing, we extract the centroid of each clustered point cloud as the UAV's point of interest, i.e., the key inspection area. This point is referred to as the POI. The data format for this point is (cluster block ID, semantic label, number of point clouds, centroid 3D coordinates, centroid 3D normal vector). 1×9 After clipping the connector neighborhood using the straight pipe section labeled 0, clustering and principal axis estimation are used to extract the central axis of each straight pipe segment. For any straight pipe segment, singular value decomposition is performed on the point cloud data matrix to extract the first principal direction. The magnitude of the projection of the pipe segment's point cloud data onto this direction is taken as the starting point of the pipe segment line. This establishes a topological graph between the straight pipe segments and connectors. (If we use Figure...) To represent the global pipeline topology, the vertex set It consists of nodes of each straight pipe section and nodes of the connector, and is an edge set. The connectivity of the inspection path is established based on the spatial proximity between the axis endpoints and the connector centers. If isolated components exist, topology repair is performed using the minimum cross-component distance to ensure connectivity of the accessible inspection path.

[0173] In the path generation layer, this invention constructs a track point sequence using the traversal node sequence in the topology graph as the skeleton, then performs smooth parameterization using piecewise conformal cubic interpolation, and shifts it by a certain set safety distance along the normal vector direction in the pipeline information, which serves as the prior path line for the UAV's subsequent model prediction contour control algorithm. To ensure that the UAV does not merely track a spatial curve, but rather forms an adjustable inspection bias within the curve's normal section, which will be discussed later... The present invention further constructs a Parallel Transport section base that propagates along the path. This method avoids the instability problem of traditional Frenet frames when the curvature is close to zero or the normal is reversed, and allows the inspection offset direction to change continuously along the entire path.

[0174] After extracting key information from the pipeline point cloud with semantic data and POI point data input from the workstation, the UAV uses a globally set coordinate system to transform the coordinates of the UAV and the actual pipeline into a global coordinate system. The UAV receives the starting or ending point of label:1 as the point with the closest calculated distance and uses it as the target starting point. The flight controller combines the pipeline point cloud and POI point position information, and determines the relative position of the target starting point through Euclidean distance calculation and an improved MPCC algorithm. The onboard computer processes data synchronously, and the LiDAR provides real-time information to correct flight parameters. When it reaches the vicinity of the target starting point, the flight controller plans the optimal path that fits the prior path and ensures safe flight based on the prior path, the virtual progress parameter θ, its own position, and the environment. The electronic speed controller precisely adjusts the motor speed according to the path plan, driving the UAV to fly at high speed and high maneuverability along the planned path.

[0175] This invention adaptively adjusts the hysteresis error weight, path advancement weight, and obstacle relaxation upper bound based on path curvature and the state of the reciprocating manifold region. This promotes a reduction in the aggressiveness of path advancement in high curvature regions, allowing the system to focus more on spatial safety and contour fit. In low curvature regions, stronger advancement is restored, improving inspection efficiency. This invention also incorporates POI inspection actions into a unified task logic. For each POI on the path, the system determines the appropriate path parameters. Sort the data and compare the current progress in real time. Relative distance to candidate POI The system selects the most recently unvisited Point of Interest (POI) that meets the conditions for a supplementary inspection window or a forward-looking window as the activation target. Upon entering POI approach mode, the next POI with a predicted path distance within a set threshold is used as the target flight position. The flight controller receives the threshold-limited deceleration signal and precisely controls the drone's deceleration. When this condition is met... or and When the system switches to hover inspection mode, it observes the POI point according to a predetermined angle sequence during the inspection. This is done by taking the average of the normals of the clusters of regions of interest near the POI point, or by calculating the normals from the relative position with the inspection trajectory. The system then fine-tunes the rotation angle of the monocular camera to take panoramic photos. After completing the required number of photos, the POI is marked as visited and the photos are transmitted back to the workstation in real time via the network. The system automatically resumes cruise and triggers a speed recovery mechanism for a certain period of time. The flight controller receives the speed-up signal and precisely controls the UAV to accelerate to the inspection speed, continuing to fly to the next POI point until the inspection is completed.

[0176] (xi) Simulation Experiment

[0177] (1) Experimental setup

[0178] Using MATLAB R2024b software, a semantic pipeline point cloud was constructed by inputting seven-dimensional data, including four semantic labels: straight pipe (label=0), endpoint (label=1), elbow (label=2), and valve (label=3). Evaluation metrics included: contour error, hysteresis error, safe distance, number of photos taken during inspection, smoothness of deceleration process, and 2D cross-sectional offset (δ1, δ2).

[0179] Visualization: Global third-person view (including pipeline point cloud, centerline, POI marker, predicted trajectory, UAV position, and safety distance line), global first-person view FPV, real-time control variables / state variables / jerk dynamic curves, MPCC error comprehensive analysis chart, POI inspection report chart, and deceleration analysis chart.

[0180] (2) Parameter settings

[0181] 1) Time and prediction parameters: Sampling time Ts = 0.05 s, prediction step size N_hor = 30, prediction duration = Ts × N_hor = 1.5 s.

[0182] 2) UAV physical parameters and safety configuration: UAV radius 0.15 m, additional safety margin 0.10 m, minimum collision clearance 0.25 m, soft constraint distance 0.40 m, hard constraint distance 0.30 m, minimum safe distance 0.25 m, emergency distance 0.15 m.

[0183] 3) Physical limitations: Maximum speed 1.5 m / s, maximum acceleration 2.0 m / s² 2 The minimum speed for inspection photography is 0.3 m / s, the maximum virtual point speed is 2.0 m / s, the virtual point speed margin is 0.6 m / s, and the virtual point θ reset distance threshold is 3.0 m.

[0184] 4) MPCC core weight parameters: contour error weight 200.0, hysteresis error weight 100.0, progress weight 30.0, acceleration smoothing weight 20.0, jerk weight 8.0, θ change rate weight 12.0.

[0185] 5) 2D cross-section offset parameters: offset back to center weight 50.0, offset smoothing weight 30.0, offset switching suppression weight 20.0, maximum cross-section offset 1.0 m, maximum offset change rate 0.5 m / s.

[0186] 6) Obstacle avoidance soft constraint parameters: obstacle relaxation variable weight 5000.0, relaxation variable squared weight 10000.0, KNN nearest neighbor query count 5, obstacle avoidance constraint steps 5.

[0187] 7) Curvature adaptive parameters: curvature factor lower limit 0.3, v_θ target velocity lower limit 0.2 m / s, curvature calculation formula 1 / (1+10κ).

[0188] 8) θ reset parameters: local search window radius 3.0 m, reciprocating manifold area reset prohibition buffer 0.5 m.

[0189] 9) SQP Iteration Parameters (Iteration Calculation): Maximum number of iterations: 2, convergence tolerance: 0.01, adaptive reduction condition: when the predicted state change is < 0.15, the iteration is reduced to 1.

[0190] 10) POI inspection parameters: hovering time 3.0 s, POI approach distance 2.5 m, POI arrival judgment distance 0.5 m, number of shooting angles 5, surround shooting radius 0.8 m, POI path association threshold 5R (5 times the pipeline radius).

[0191] 11) Reciprocating manifold area parameters: area deceleration factor 0.3.

[0192] (3) Simulation results

[0193] I) Path tracking accuracy

[0194] like Figures 1-4 As shown, the contour error remains within a small range most of the time, indicating that the UAV can accurately track the cross-sectional direction of the reference path. The hysteresis error increases slightly in the curved pipe section and the reciprocating manifold region, but it can converge quickly through curvature adaptive speed adjustment and θ local reset mechanism. The MPCC of this invention decouples the two and optimizes them uniformly through the coupling matrix, avoiding the conflict between tracking accuracy and progress control.

[0195] II) Obstacle Avoidance Safety

[0196] The distance between the drone and the pipe surface was consistently higher than the minimum safe distance (0.25m), fluctuating only slightly within the relaxation boundary allowed by QP (an adaptive relaxation variable). Figure 2 (D) Figure 4 (D) KD-tree speeds up a single nearest neighbor query to O(log n) complexity compared to brute-force search, ensuring real-time performance (e.g., Figure 5 The external emergency obstacle avoidance layer serves as a last resort in cases of QP solution failure or extreme proximity. Figure 2 (D) Figure 4 (D)).

[0197] III) Section offset control

[0198] The offsets δ1 and δ2 change smoothly within the range of [-1, 1]m. The offset is mainly used to ensure that after local obstacle avoidance, the system can return to 0 through weighted regression (i.e., follow the centerline). Figure 2 (C) Figure 4 (C)). The offset smoothing weight and rate of change constraint ensure the continuity of the offset motion and avoid sudden jumps.

[0199] IV) POI Inspection Performance

[0200] The system can automatically identify Points of Interest (POIs) along the path and dynamically select the nearest unvisited POI based on distance. Figures 6-9 It can be seen that the drone decelerates smoothly in advance before arrival, and the deceleration process upon arrival is smooth (the deceleration time and speed curve can be quantitatively analyzed).

[0201] After inspecting a POI, the drone quickly returns to cruise mode after an 8-second speed recovery period. If the next unvisited POI is within the deceleration range during the 8-second recovery period, deceleration logic is prioritized. Figures 8-9 It can be seen that the drone smoothly decelerated and accelerated, ensuring inspection efficiency. It can also be seen that multiple decelerations occurred in the reciprocating manifold area of ​​complex pipelines to match the speed and safety of this stage.

[0202] V) Adaptability to complex pipelines

[0203] Depend on Figure 8 In a simple pipe, the high curvature area can be identified within 60-75 seconds. This corresponds to a pre-emptive deceleration at the pipe bend, followed by a speed-up after the image is taken, consistent with basic experimental and algorithm settings. A subsequent deceleration occurs within 65-70 seconds, corresponding to crossing the pipe and hugging its wall. Emergency obstacle avoidance ensures a smooth crossing without touching the pipe, allowing the inspection to continue, thus fulfilling the experimental settings. Figure 9 It can be seen that the drone encountered multiple high curvatures in the reciprocating manifold area between 80 and 120 seconds. From the moment it entered, it automatically reduced its speed to 0.3 times its initial speed and remained at a low speed. The high curvature bends reduced the v_θ speed through curvature adaptive reduction, ensuring that the drone could conduct stable sequential inspections and would not miss the next POI due to excessive lag error.

[0204] Because the reciprocating pipe section in the reciprocating manifold area is prone to drone lag due to obstacle avoidance and photography, if the v_θ speed is not controlled, the virtual pathpoint θ may circle back and forth before directly selecting the virtual pathpoint's position to follow, instead of following the prior pathline. This invention sets the v_θ speed based on the lag error (the larger the lag error, the lower v_θ), effectively resetting the virtual point and ensuring subsequent inspections (…). Figure 3This verifies that the flight trajectory almost follows the prior path. To avoid obstacles, obstacle avoidance constraints are dynamically relaxed in high-curvature or reciprocating manifold regions. Figure 4 (D) It can be seen that the drone can get further away from the obstacle at this stage, ensuring safety and inspection efficiency.

Claims

1. A UAV pipeline inspection control method based on model prediction profile algorithm, characterized in that, It includes the following steps: (1) By processing the input pipeline point cloud data with semantic labels, coordinate units are unified and point sets are classified; For single-point data Perform a unified unit conversion to obtain the set of straight pipe points and the set of non-straight pipe points; This represents the three-dimensional coordinates of the pipe relative to the global coordinate system. The label represents the normal vector value for the regional pipeline cluster after subsequent calculations. i For semantic tags, Indicates a straight pipe. This indicates the two ends of the pipeline, i.e., the start and end points. Indicates a bend. Indicates a valve; (2) Cluster the endpoints, elbows, and valve points to obtain the connector center and POI candidates; construct the straight pipe center axis, and establish a topology graph with straight pipe sections and connectors as nodes; the POI candidates are the pipeline point cloud after removing... Points outside the straight pipe; (3) Generate discrete inspection track points based on the topology map, and generate a safety inspection trajectory by combining the offset mode; Gaussian smoothing and PCHIP interpolation are performed on discrete trajectory points to obtain the path function. and its derivative and calculate the curvature; (4) Construct the Parallel Transport section base along the path ; Calculate the nearest distance and corresponding parameters to all POIs, filter the valid POIs for the path, and sort them. (5) Establish an extended state model; in each sampling period, based on the current state, curvature, reciprocating manifold area, POI state and obstacle distance, solve the quadratic programming online and output the acceleration, path advancement speed and cross-sectional offset change rate; The extended state model is a discrete linear extended state model, and the state and control quantity at the k-th sampling time are respectively represented as follows: ， in, For drone three-axis acceleration, For the path progress evolution speed, For the pipe section offset velocity, This represents the two-dimensional offset within the cross-section of the UAV's pipe. Discrete systems , , , and These are the three-dimensional and two-dimensional identity matrices, respectively. The update process corresponding to the extended state model is as follows: ; (6) When the UAV approaches the POI, it enters the approach mode and enters the hovering inspection mode when the criterion is reached to perform multi-view photography; when all inspection tasks are completed, the UAV reaches the destination and performs the ascent and hovering end action.

2. The UAV pipeline inspection control method based on model prediction contour algorithm according to claim 1, characterized in that, In step (2), the specific method for establishing the topology graph is as follows: (a) For any straight pipe segment, establish its axis start and end points by extracting the main axis. And construct the axis: Where τ is a random point parameter of the pipe segment axis; when τ is 0, it is the starting point, and when it is 1, it is the ending point. (b) Establish a diagram structure based on each straight pipe section and each connector. ; If the endpoint of the straight pipe section and the center of a certain connector satisfy the following... , Therefore, edges are added between the corresponding nodes. This is the topology connectivity threshold; The center of a certain connector; if there are pipe segment nodes in the diagram with a degree less than the requirement, topology repair is performed according to the minimum distance criterion; If multiple connected components exist, then for any two components Select ;in This represents the spatial location of a graph node; adding connecting edges to this node yields a globally traversable topology.

3. The UAV pipeline inspection control method based on model prediction contour algorithm according to claim 2, characterized in that, In step (3), the formula for calculating the curvature is: in, r(θ) is the path function; set up The final path function is , , They are respectively , , Obtained after performing PCHIP interpolation; For discrete points of the trajectory, This is the cumulative arc length.

4. The UAV pipeline inspection control method based on model prediction contour algorithm according to claim 3, characterized in that, In step (4), the Parallel Transport section base The construction method is as follows: along the discrete parameter sequence Calculate the tangent vector in the direction At the initial point, if the unit vector of the pipeline is... If the reference axis is not parallel to the world vertical axis, select the world vertical axis as the initial reference; otherwise, select the world horizontal axis as the reference. Let the initial reference axis be... The initial normal basis is defined as , For subsequent discrete points, if the change in adjacent tangential directions is significant, propagation is performed according to the rotation formula; if the change is minor, the normal basis from the previous step is directly maintained; finally, for any continuous parameter... The result is obtained by linear interpolation and normalization of the nearest discrete basis vectors. , .

5. The UAV pipeline inspection control method based on model prediction contour algorithm according to claim 4, characterized in that, In step (4), the method for filtering and sorting valid POIs along the path is as follows: all connector centroids and external input POIs are uniformly represented as , Indicates the POI location. For POI normal direction, Define the semantic labels for POIs; calculate the minimum distance from each POI to the path: Only when satisfied Only then is the POI considered a valid POI for the path, and its corresponding path parameter is recorded as... , The path association threshold; all valid POIs are set according to... After sorting, a sequence of key points for scheduling inspection tasks is formed.

6. The UAV pipeline inspection control method based on model prediction contour algorithm according to claim 5, characterized in that, In step (5), the error cost of solving the quadratic programming problem at step j is: in, for The transpose of ; W j For a comprehensive weighted matrix, in, For the contour error weights of a single step, P represents the weight of the single-step lag error. c and P l It is the projection matrix of contour error and hysteresis error; Let j be the error vector at step j. Among them, the error state coefficient matrix , for The state at time t, the number of remaining parameters after removing the state variables. ; For path parameters, path points tangent vector of path point The path point corresponds to the cross section base. .

7. The UAV pipeline inspection control method based on model prediction contour algorithm according to claim 6, characterized in that, In step (5), the cost function of the extended state model is: in, To track error terms, To control input penalty items, For fluid-level control smoothing terms, For reference, track progress items, For cross-sectional regression terms, This is the relaxation term of the mixed-line quadratic form.

8. The UAV pipeline inspection control method based on model prediction contour algorithm according to claim 7, characterized in that, In step (6), after entering the approximation mode, the mode velocity factor is set to... To approximate the distance threshold for POI, Minimum speed factor For the progress of the i-th POI Compared with the current progress The distance between them; The arrival criterion is satisfied At that time, or when satisfied and When this happens, it switches to hovering inspection mode; among which, The progress difference threshold, For POI target points, Reaching the distance threshold, Speed ​​threshold.

Images